Understand AI.
From first aha
to better judgment.

Generative AI produces text, images or other content from patterns learned during training. For a language model, producing an answer involves successive predictions of tokens: words, word pieces or symbols.

Context includes the instructions and information available when an answer is generated. Describing the audience and task changes what the model may propose. That is not a fact-check.

Confident wording proves neither human understanding nor accuracy. The tool can help explore or rephrase ideas; a person still needs to judge the result for its intended use.

A next token receives a probability based on context. The answer is built one piece at a time. A natural-sounding continuation can include invented information.

Temperature changes the distribution used for sampling. In this simulation, a low temperature concentrates possibilities; a higher temperature spreads them out. This setting does not measure truth.

Comparing two outputs can reveal variation. To know which is accurate, you need independent evidence: a source document, an observation or a check suited to the task.

A summary can add details absent from its source. Before sharing, separate the claims: who, what, when, how much, and supported by what evidence?

Supported means the source backs the claim. Contradicted means it says something incompatible. Not stated means it does not let you conclude. Missing information is not automatically proof of the opposite.

A reference supplied by a tool also needs checking: does it exist, does it say what is attributed to it, and is it relevant? A well-formatted citation can be incorrect.

A useful instruction specifies the desired result, audience and context. “Make it better” leaves much implicit; “prepare three examples for beginners” offers a clearer starting point.

Add a format and criteria: length, language, separation of facts and assumptions, and items to verify. Asking the tool to state its limits may help the reader, but those statements still need examination.

The first answer is a draft to evaluate. Compare it against your criteria, identify a specific problem, then refine the instruction. A better instruction makes evaluation easier; it does not guarantee accuracy.

Before sending text to a service, ask what it reveals. A name, address, learning difficulty or customer record may concern a person. An internal project may also contain confidential information.

Removing a name is not always enough. Age, location, role and an unusual situation may identify someone. For learning or demonstration, a fully fictional example often avoids transmitting the real document.

The right action also depends on the tool and your organization’s rules: approved access, settings, retention and responsible people. If unsure, pause and ask for guidance. This activity does not replace those rules.

Evaluating a tool with one number can conceal differences between groups, languages or situations. An overall result does not necessarily describe each person’s experience.

In our fictional example, success is fixed at 95% for group A and 60% for group B. The slider changes only the sample composition. It does not train a model or change either group’s results.

A higher average may simply mean the better-served group is more common in the test. Before deciding, examine the relevant situations, who is underrepresented and the consequences of an error.

You are organizing a workshop. A tool proposes a convincing announcement from your notes. Your job is to decide what to keep, what to correct and what must not be transmitted.

Use a simple method: define the intention, limit data, specify criteria, verify claims and identify who approves the result. If an error would have significant consequences, increase the level of checking or choose another method.

After the assessment, take the practice sheet. Choose a small task, define what you will check and compare the result with your usual method. A useful experiment may also conclude that AI does not help with that task.